PhD Thesis and MSc Reports

Author

Download

Modelling Visual Search for Surface Defects
Ph.D. thesis, School of Mathematical and Computer Sciences, Heriot-Watt University, 2010.

Abstract
Much work has been done on developing algorithms for automated surface defect detection. However, comparisons between these models and human perception are rarely carried out. This thesis aims to investigate how well human observers can nd defects in textured surfaces, over a wide range of task diculties. Stimuli for experiments will be generated using texture synthesis methods and human search strategies will be captured by use of an eye tracker. Two dierent modelling approaches will be explored. A computational LNL-based model will be developed and compared to human performance in terms of the number of xations required to nd the target. Secondly, a stochastic simulation, based on empirical distributions of saccades, will be compared to human search strategies.

A Psychophysically-based Model for the Perceived Directionality of Textured Surfaces
Ph.D. Thesis, School of Mathematical and Computer Sciences, Heriot-Watt University, 2010.

Abstract

Directionality is known to be an important dimension in human perception and classification of visual textures. Through a series of psychophysical experiments, this thesis investigates further the perception of directionality in textured surfaces, and uses the results to propose a measurement model for the perceived directionality of random-phase surfaces.

Height maps of textured surfaces were rendered and animated in real-time with controlled illumination. Observers' judgements of the directionality of surfaces were obtained by direct-ratio estimation, and either the method of pair-wise comparisons or the method of constant stimuli. The responses were used to derive a perceptual scale of directionality (perceived directionality) that could be related to physical properties of the surfaces.

The thesis first investigates the relationships between each of two existing computational measures of directionality (Tamura's variance and Davis' variance) and human perception of directionality. This was done by using height maps captured from real surfaces, which were then manipulated to vary the computational measures of their directionality. From the psychophysical experiment, it was found that these two measures do not fully account for human perception of directionality, which must therefore be influenced by other properties of the textures.

In order to investigate more fully the factors determining perceived directionality, synthetic random-phase surfaces defined by a mathematical model were used in the subsequent experiments. It was found that three properties of the magnitude spectrum of such surfaces significantly affect human perception of their directionality: angular variance, RMS roughness and central radial frequency. After determining these effects, the thesis proposes a measurement model of perceived directionality, which predicts human perception of directionality of a random-phase surface.

Perceptual Dimensions for Surface Texture Retrieval


Ph.D. Thesis, School of Mathematical and Computer Sciences, Heriot-Watt University, 2008.

Abstract

This thesis presents a methodology for developing perceptually relevant surface texture retrieval systems. Generally such systems have been researched using image texture which has been captured under unknown or uncontrolled conditions (e.g. Brodatz). However, it is known that changes in illumination affect both the visual appearance of surfaces and the computational features extracted from their images. In contrast this thesis either uses surface information directly, or computes features obtained from images captured under controlled lighting conditions.

Psychophysical experiments were conducted in which observers were asked to place texture samples into groups. Multidimensional Scaling was applied to the resulting similarity matrices to obtain a more manageable reduced perceptual space. A four dimensional representation was found to capture the majority of the variability. A corresponding feature space was created by linearly combining selected trace transform features. Retrieval was performed simply by determining the n closest neighbours to the query’s feature vector. An average retrieval precision of 60% was obtained in blind tests.

Due to some confidentiality agreement, the content of this thesis will be made available at a later time.

Mathematical Models for Perceived Roughness of Three-Dimensional Surface Textures
Ph.D. Thesis, Heriot-Watt University, 2008.

Abstract

This thesis reports and discusses results from a new methodology for investigating the visually perceived properties of surfaces; by doing so, it also discovers a measurement or estimator for perceived roughness of 1 / F ^&beta noise surfaces.

Advanced computer graphics were used to model natural looking surfaces (1 / F ^&beta noise surfaces). These were generated and animated in real-time to enable observers to manipulate dynamically the parameters of the rendered surfaces. A method of adjustment was then employed to investigate the effects of changing the parameters on perceived roughness. From psychophysical experiments, it was found that the two most important parameters related to perceived roughness were the magnitude roll-off factor (&beta) and RMS height (&sigma) for this kind of surfaces.

From the results of various extra experiments, an estimation method for perceived roughness was developed; this was inspired by common frequency-channel models. The final optimized model or estimator for perceived roughness in 1 / F ^&beta noise surfaces found was based on a FRF model. In this estimator, the first filter has a shape similar to a gaussian function and the RF part is a simple variance estimator. By comparing the results of the estimator with the observed data, it is possible to conclude that the estimator accurately represents perceived roughness for 1 / F ^&beta noise surfaces.

Table of contents and individual chapters are available here.

Three-dimensional Surface Texture Synthesis
Ph.D. Thesis, Heriot-Watt University, 2003.

Abstract
Texture synthesis has been extensively investigated by both computer vision and computer graphics communities during the past twenty years. However, the input and output are normally 2D intensity texture images. If the subjects are 3D surface textures (such as brick, woven or knitted textiles, embossed wallpapers etc.), these 2D synthesis techniques cannot provide the information required for rendering under other than the original illumination and viewpoint conditions. The aim of this thesis therefore is to develop inexpensive approaches for the synthesis of 3D surface textures. Few publications are available in this research area.

We first introduce an overall framework for the synthesis of 3D surface textures. The framework essentially combines surface representation methods with 2D texture synthesis algorithms to synthesise and relight new surface representations. Then we investigate five low-dimensional methods, namely the 3I, Gradient, PTM, Eigen3 and Eigen6 methods, for extracting representations from a set of images of the 3D surface texture sample. The surface representations can be relit to generate new images under arbitrary lighting directions by linear combinations. These methods are quantitatively assessed by comparing the original and relit images. The results show that the Eigen6 produces the best performance.

We select a 2D texture synthesis algorithm which is then extended into multi-dimensional space to use the five surface representations as input. In this way, we develop five approaches for the synthesis of 3D surface textures. The synthesised results are compatible with computer graphics systems and can be used in real-time rendering applications. The five synthesis approaches are qualitatively assessed by employing psychophysical experiments and non-parametric statistics. The results show that the two low-dimensional methods, the Gradient and Eigen3, on average offer as good a performance as of any of the other methods and incur low computational cost.

Table of contents and individual chapters are available here.

Height Recovery of Rough Surfaces from Intensity Images
Ph.D. Thesis, Heriot-Watt University, February 2003.

Abstract
This thesis is concerned with the 3D estimation of rough surfaces from their intensity images. A technique which combines Photometric Stereo and frequency integration is proposed. The combination of these two standard methods for reconstructing rough surfaces is novel. We refer to this technique as the Benchmark technique. Two novel recovery algorithms which rely on assumptions about the linearity of the surface reflectance are also presented. We refer to them as the Optimal Linear Filter and the Linear Photometric Stereo. The proposed methods differ in the information that they require as well as in the assumptions that they make about the surface.

The ability of the proposed techniques to estimate rough surfaces is assessed using simulation and real data. The assessment considers a diverse set of textures including those that are challenging for the algorithms, such as very rough or specular surfaces.

The most robust estimation is given by the Optimal Linear Filter. However this technique requires information about the surface topography, which is usually not available. Between the alternatives, the Benchmark technique gives more accurate reconstructions.

A post-processing step which can be used to improve the surface estimate is presented. This minimises the brightness error using an iterative approach. When the Linear Photometric Stereo method is combined with the post-processing step, its performance is similar to that of the Benchmark technique, despite requiring one less image. However the Linear Photometric Stereo algorithm is restricted to constant albedo surfaces. The choice of the most appropriate method is determined by the application requirements.

Table of contents and individual chapters are available here.

Rotation Invariant Classification of 3D Surface Texture Using Photometric Stereo
Ph.D. Thesis, Heriot Watt University 2003.

Abstract
This thesis presented a new three-dimensional surface texture classification scheme which was invariant to surface-rotation using photometric stereo. Many texture class ification approaches had been presented in the past that were image-rotation invariant, however, image rotation was not necessarily the same as surface rotation. A classifier therefore had been developed that used invariants that were derived from surface properties rather than image properties.

Firstly, various surface models were considered and a classification scheme was developed that used magnitude spectra of the partial derivatives of the surface obtained using photometric stereo. A simple frequency domain method of removing the directional artefacts of partial derivatives was presented, and a 1D feature set of polar spectrum was also extracted from resulting spectrum. Classification was performed by comparing training and classification polar spectra over a range of rotations. Secondly, a new feature generator albedo spectrum was introduced to provide more information on surface texture properties, and an additional 1D feature set of the radial spectrum was employed too. In addition, by examining the effect of shadowing, a four-image photometric stereo method was developed to provide more accurate three-dimensional surface properties. Finally, a verification step was included in the classification where the 2D spectrum features were compared instead of 1D spectrum features.

The classification results using new-developed photometric stereo real texture database shown that combining 2D gradient and albedo data improves the classification's performance to provide a successful classification rate of 99%

Table of contents and individual chapters are available here.

The Classification of Textured Surfaces Under Varying Illuminant Direction
Ph.D. thesis, Heriot-Watt University 1998.

Abstract
This thesis sets texture analysis in a physical context. Models of the system components are obtained from the literature and integrated into a description of the process linking the rough surface to the feature set on which classification is based. The first component is the rough surface, models of the surface topography are selected from the fields of tribology and scattering. Various reflectance models are considered and a spectral model of the surface/image relationship from the literature, is evaluated and discussed. The relationship between the incident image and the captured data set is investigated and described. This model is integrated with the spectral description of the feature measures to form a model of the transition from surface to feature set.

It is clear from this model that the direction of illumination can affect the directionality of an image obtained from a given surface. Changes in the illuminant direction will result in changes in the feature outputs. If the illuminant direction is altered between training and classification, the classification rule may be inappropriate and classification poor. Several schemes are considered to combat this problem. A technique which uses a representation of the physical surface as the basis for the generation of appropriate training data is selected for further evaluation. The surface derivative fields of the training surface are estimated using photometric techniques. A rendering algorithm uses these estimates to simulate the appearance of the training surface when it is illuminated from an arbitrary direction. It is shown that where illuminant direction is varied this system is able to perform significantly better than a naive classifier, and in some cases approaches the level of accuracy obtained from training the classifier under the conditions at which classification is performed.

Table of contents and individual chapters are available here.

The Effect of Varying Illuminant Direction on Texture Classification
Ph.D. thesis, Dept. Computing and Electrical Engineering, Heriot-Watt University, August 1994.

Abstract
Texture analysis has been an extremely active and fruitful area of research over the past twenty years. Many advances have been made, but the effect of variation in lighting conditions on automated texture classification and segmentation has received little attention. This thesis shows that the direction of the illuminant is an important factor that should be taken into account when analysing images of three-dimensional texture.

A frequency domain model is presented which predicts that both the directional characteristics and the variance of images of three-dimensional texture can be affected by changes in illuminant vector. Results of simulations and laboratory experiments support these predictions.

The responses of three sets of texture measures are analysed using a test set of isotropic and directional textures. The results show that the feature measures' outputs are affected by changes in illuminant direction. These changes are also shown to significantly increase the error rates of statistical classifiers implemented using the three feature sets. Normalisation of images is shown to reduce the error rates in some cases.

The frequency domain model of image texture is further developed using empirical data and the resulting model used to design a set of tilt-compensation filters. These filters are used to pre-process images to reduce the effects of changes in the angle of tilt of the illuminant. Application of the filters to the test image set reduced the classification errors associated with directional textures.

Table of contents and individual chapters are available here.

Author

Download

I.S. Bothwick

The Effect of Ensonification direction on seabed Images and Classification
MSc Thesis, September 1998, Dept. of Computing and Electrical Engineering

Novel image processing of 3D textures
MSc Thesis, Heriot Watt University, September 2003

Abstract
A new invariant-rotation texture operator, known as LBPROT (Local Binary Pattern Rotation-Invariant), has been recently developed by M. Pietikäinen, T. Ojala and Z. Xu. It has demonstrated much better performance at classifying textures than the well-known CSAR (Circular-Symmetric Autoregressive Random Field). This paper extends the experiments carried out then, and boards an alternative series of experiments in order to find out further information regarding LBPROT operator's behaviour.

Among the experiments performed, an analysis of the operator's variability before distinct samples of the same texture under equal illumination conditions was accomplished. Furthermore, a research aiming at understanding the operator's response when applied to different directionality features is widely presented. Moreover, some extra experiments utilize the operator output distribution to classify textures by using the G Statistic log-likelihood pseudo-metric. Finally, all these investigations are assessed leading to a series of interesting results which are discussed in dept.

Uncalibrated Photometric Stereo for 3D Surface Texture Recovery
Research Memorandum RM/02/02, May 2003, School of Mathematical & Computer Sciences

Abstract
This dissertation presents the method of uncalibrated photometric stereo for estimating the surface normal and the reflectance field without a priori knowledge of the light-source direction or the light-source intensity. 

In this method, assuming only that the object's surface is Lambertian, the surface normal, and the surface reflectance, the light-source direction, and the light-source intensity can be determined simultaneously.